\u300aCoal Gangue Sorting Dataset with Imagery, Virtual Configuration, Programs and Logs for Detection & Recognition Tasks\u300b

Coal-Gangue sorting is a key part of safe mining production. The complex geological conditions of coal mining areas, along with the diverse forms and mineral components of the mined materials, pose significant challenges to the further development of Coal-Gangue sorting processes and methods. Traditional manual or semi-manual sorting methods face harsh working conditions and low efficiency. Machine vision technologies and models, have issues like model complexity, large training sample requirements, and poor generalization ability. This leads to high annotation costs and poor sorting performance when applying machine vision to Coal-Gangue sorting. To address these problems, we propose Coal-SAM, a novel detection-segmentation paradigm that combines efficient detection tools with the precision of the Segment Anything Model (SAM). By using detector-generated bounding boxes to guide SAM-based segmentation, this framework enables automated annotation with minimal manual intervention, reduces labor costs, and improves Coal-Gangue sorting accuracy. We establish CGSD-SP, the first specialized Coal-Gangue dataset featuring professionally annotated multi-scenario images to facilitate adaptive model training. Experimental validation confirms Coal-SAM's industrial-grade performance in semantic segmentation, fulfilling practical requirements through three technical innovations: (1) geometric consistency optimization via cooperative parameter alignment, (2) a cost-effective end-to-end annotation workflow, and (3) a dedicated inference engine for dynamic environments. The framework demonstrates robust generalization in complex industrial settings, substantiated by the systematic verification of CGSD-SP's effectiveness in model optimization, thereby providing crucial technical support for the intelligent advancement of Coal-Gangue sorting.Gangue Detection-Robotic Control Architecture (GDRCA) is a collaborative framework in the coal industry. It enables robotic arms to automatically sort coal, gangue, and debris on high-speed conveyors, thereby improving production efficiency and safety. Its key breakthroughs focus on improving adaptation to conveyor speed, load handling capacity, and sorting accuracy. This study optimizes the Enhanced GDRCA(EGDRCA) robotic arm control architecture by integrating big data and image recognition technologies, proposing three innovations: EGDRCA achieving smooth power transition through the collaboration of pre-imaging processing and robotic arm pre-tracking, and developing a power consumption-constrained learning loss algorithm to quantify the adjustment effect of high-power pulses; constructing a EGDRCA-Collaborative Sample Generator (EGDRCA-CSG) that orchestrates a multi-model collaboration among target detection, semantic segmentation, and sample generation modules\u2014by leveraging detection masks as prompts and fine-tuning the sample creator with only 5% annotated data, the robotic arm maintains a high recognition rate (mAP50 over 73%) for grasping tasks in low-annotation scenarios; proposing the Spatio-Temporal Power-Aware Non-Maximum Suppression (ST-PA NMS) algorithm to address the issue of repeated detection of dynamic targets in high-speed conveyor scenarios, by enhancing the traditional NMS through integrating a spatio-temporal continuity score into its ranking mechanism (based on target motion trajectory prediction) and a power consumption penalty term (positively correlated with the real-time current of the robotic arm). Experiments show that the EGDRCA, via the power consumption-constrained learning loss algorithm, reduces the current pulse peak-to-valley difference by 30.4%, with a laboratory simulation success rate(Succ_rate) of 94.7%; the sample generator maintains a mean Average Precision at 50% IoU (mAP50) of over 73% with only 5% annotated data.